02. Tool Calling & Structured Output

Overview

In this session, we evolve from regex-based parsing to structured tool calling using Pydantic and OpenAI's Function Calling API. This gives us type safety, validation, and reliability.

Why Function Calling?

Key Concepts

Pydantic for Tool Schemas

Define your tool parameters with full type safety:

from pydantic import BaseModel, Field
 
class CalculatorInput(BaseModel):
    """Input schema for calculator tool"""
    expression: str = Field(
        description="Mathematical expression to evaluate"
    )

Converting to OpenAI Format

OpenAI expects a specific JSON schema format:

def pydantic_to_openai_function(model, name, description):
    return {
        "type": "function",
        "function": {
            "name": name,
            "description": description,
            "parameters": model.model_json_schema()
        }
    }

The Function Calling Flow

Hands-on Practice

In the notebook, you will:

1. Define Pydantic Schemas - Calculator, Search, Weather tools
2. Convert to OpenAI Format - Build the tools array
3. Implement Tool Handlers - Actual execution logic
4. Build FunctionCallingAgent - Complete agent class
5. Integrate Tavily - Real web search capability

Tavily Integration

from tavily import TavilyClient
client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
results = client.search(query, max_results=5)

Comparison: Before vs After

Before (Regex Parsing)

# Fragile - breaks easily
pattern = r"Action:\s*(\w+)\[(.+?)\]"
match = re.search(pattern, response)

After (Function Calling)

# Robust - guaranteed structure
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=messages,
    tools=tools,
    tool_choice="auto"
)
 
for tool_call in response.message.tool_calls:
    # Type-safe, validated arguments
    args = json.loads(tool_call.function.arguments)

Best Practices

1. Descriptive Schemas - Better descriptions = better tool selection
2. Error Handling - Wrap tool execution in try/except
3. Logging - Log all tool calls for debugging
4. Timeouts - External APIs can hang

Next Steps

Ready to build something real? Head to the Weekend Project to create a Personal Research Assistant!